Our ability to make reliable physical predictions even in novel settings is a hallmark of human intelligence. Here we investigate how people infer multiple physical variables simultaneously and compose them to generalize to a novel scenario. Participants (N=203) observed a series of balls launched at different angles in a 2D virtual environment and generated predictions about their trajectories. We found that people could infer the masses of different balls based on these observations, as well as the existence of a latent "wind" force, and compose knowledge of these two variables to generalize to novel situations in a subsequent test phase. We modeled this generalization as the consequence of being able to simulate trajectories by independently combining force and mass information in accordance with Newtonian mechanics. To validate this approach, we also tested several alternative models and compared their generalization behavior to one another and to that of people. Together, our study points to the value of using generalization to probe the underlying representations supporting physical prediction.